Exploration on the Application of Neural Network Technology in the Protection and Inheritance of Ethnic Traditional Sports Culture

When using multi-node sensor units for human motion capture, initial alignment of the human body is required for the accuracy of the system. Since the muscles of the human body are prone to deformation, there is no way to ensure that the sensor unit is oriented in the same direction as the limb when wearing the sensor unit, and alignment is required. In addition, when the human body is reconstructed, it is also necessary to unify the posture of each part, and the unified coordinate system should be consistent with the geographic coordinate system. Therefore, the initial alignment is mainly between the reference coordinate system and the navigation coordinate system.

Effective alignment of the coordinate systems of the sensor units and the parts of the limbs is achieved by reasonably wearing the sensor units, with four sensor units being worn along the hand on the outer side of the large arm and the small arm, and four sensor units being worn along the leg on the outer side of the large leg and the small leg. The coordinate systems of the left big arm, the left small arm, the left big leg, the left small leg, the right big arm, the right small arm, the right big leg, and the right small leg correspond to the coordinate systems of the eight sensor units, respectively. When worn, the correspondence between the sensor units and the coordinate systems of each limb is realized according to the arrows on the sensor unit housing as a reference. After wearing, the coordinate system of each node sensor unit and the corresponding limb coordinate system coincide. Therefore, the data of the node is the data of the corresponding limb.

The hardware composition of the national traditional sports motion capture technology based on inertial sensors is shown in Figure 2. This system selects ICM208 nine-axis inertial sensor to realize the acquisition of human posture data, which integrates three-axis accelerometer, three-axis gyroscope and three-axis magnetometer. In order to solve the problem that the traditional 17 joints of human body motion capture system can not collect the fine movements of the hand, this paper designs a motion capture glove based on the inertial sensor-based national traditional sports motion capture technology. In the traditional motion capture technology only in the back of the hand to install a sensor on the basis of the new 20 sensors, that is, each finger to install four sensors, a total of 21 sensors installed in a motion capture glove, both hands plus other parts of the human body 15 sensors installed a total of 57 nine-axis inertial sensors.

Figure 2.

Based on the inertial sensor’s action capture hardware

The data transmission equipment mainly includes XC7Z02 embedded system, Wi-Fi module, router, etc. XC7Z02 embedded system adopts ARM+FPGA SOC technology to integrate the dual-core ARM Cortex-A9 and field-programmable gate array (FPGA) on a single chip, which reduces the size of the system and improves the data interaction rate between the micro-control unit (MCU) and the FPGA. Interaction rate between the MCU and FPGA. In order to reduce the signal interference between high-speed data lines and improve the quality of data transmission, shielded wires are used to interface between the 57 inertial sensors and between the inertial sensors and the FPGA. The connection between each SPI line and the SOC module is realized by plugging and unplugging interfaces, and measures such as buffering the SPI signals are taken to avoid problems such as signal reflection and ringing.

The complementary filtering algorithm is applicable to the scenario where different types of sensors measure the same physical quantity and their corresponding spectra differ due to the different characteristics of different measurement devices [13]. The signal acquired by one sensor device contains the real physical quantity signal and low-frequency noise, while the signal acquired by the other sensor device contains the real physical quantity and high-frequency noise, and the low-frequency noise can be decomposed by subtracting the two measurement signals and passing them through a low-pass filter, which can optimize the state estimation.

After obtaining the data of the national traditional sports movements using inertial sensors, it is necessary to settle the attitude of the sports movements, and there are two methods to calculate the attitude of the nine-axis IMU, one is to solve the attitude by integrating the angular velocity, and the other is to solve the attitude by orthogonal decomposition of the linear acceleration. However, since both methods will have result bias and high frequency noise, so we can utilize complementary filtering to fuse two kinds of data for attitude solution, for nine-axis IMU complementary filtering is to compensate the gyroscope by accelerometer.

Set the acceleration of gravity g in the geographic coordinate system and transform g from the geographic coordinate system to the carrier coordinate system using the coordinate transformation matrix to obtain the theoretical acceleration of gravity g_b in the carrier coordinate system. i.e: (1)gb=CRb⋅g=(CbR)−1⋅[ 0 0 1]=[ 2(q1q3−q0q2) 2(q2q3+q0q1) 1−2(q12+q22)]=[ gbx gby gbx]

From Eq. It can be seen that the gravity acceleration is exactly the last column of the matrix after transforming from the geographic coordinate system to the carrier coordinate system, thus the theoretical gravity acceleration g_b derived from the quaternion describing the attitude of the carrier is obtained, and in addition the actual gravity acceleration g_t can be measured by using the accelerometer. g_b and g_t are vectors mathematically, and the deviation can be expressed in terms of the angle relationship between the two. Considering that the vector outer product mode ui vector pinch angle is positively correlated, the vector direction difference θ is obtained by calculating the outer product. That is: (2)|ρ|=|gb|⋅|gt|⋅sinθ

Before performing the outer product calculation, g_b and g_t should be unitized, so there: (3)|ρ|=sinθ

Considering that usually g_b does not deviate from g_t by more than 45°, and that sin θ is approximately equal to θ when θ is within ±45°, the above equation can be simplified to: (4)|ρ|=θ

After obtaining the vector deviation, the PI compensator is constructed to calculate the angular compensation value as: (5)GyroError=Kp⋅|ρ|+KI⋅∫|ρ|

The proportional term is used to control the “confidence” of the sensor and the integral term is used to eliminate static errors. The larger K_P is, the more significant the effect of error compensation through the accelerometer is, and conversely, the smaller K_P is, the weaker the accelerometer’s compensation of the gyroscope is, and the data is more dependent on the gyroscope data. The integral term is used to eliminate biased noise in the angular velocity measurements, so for the angular velocity measurements corrected for zero deviation, a very small K_I is generally chosen, and finally the angular velocity measurements are added to the compensated value and brought into the quaternion difference equation to update the current quaternion.

A Hidden Markov Model (HMM) is a stochastic process with a double embedded structure of Markov chains, where one is a sequence of state observations and the other is an implied sequence of state transfers and is in a finite sequence of a certain class, whose state transition process is unobservable [14].

The evaluation problem of HMM is solved utilizing forward and backward algorithmic solutions, where the forward variables are computed by the recurrence relation defined by the forward algorithm Eq. (6), and the backward variables are computed by the recurrence relation defined by the backward algorithm Eq. (7), and the probability of occurrence of the observed sequences can be found by Eq. (8). That is: (6)α(t+1,j)=∑i=1Nα(t,i)aijbj(Oi+1) (7)β(t,i)=∑j=1Nβ(t+1,j)aijbj(Oi+1) (8)P(o1,⋯,oT)=∑i=1Nα(T,i)

The decoding problem for HMM is usually defined using the Viterbi algorithm, viz: (9)δ(t,i)=maxP(q1,⋯qt−1,qt=Si,o1,⋯,ot|λ)

The learning training problem of HMM is solved by the Baum-Welch algorithm, which calculates the current maximum likelihood estimate of the model to solve the model training. Then: (10)ξt(i,j) = P(qt=Si,qt=1=Sj|O,λ) = α(t,i)aijbjβ(t+1,j)P(O|λ) (11)γt(i)=P(Si|O,λ)=α(t,i)β(t,i)P(O|λ) (12)ξt(i,j) = P(qt=Si,qt+1=Sj|O,λ) = P(qt=St,qt+1=Sj,O|λ)P(O|λ) = αt(i)aijbj(Ot+1)βt+1(j)P(O|λ) = αt(i)aijbj(Ot+1)βt+1(j)∑i=1N∑j=1Nαt(i)aijbj(Ot+1)βt+1(j)

HMM model is a stochastic process with double embedded structure, which has strong timing modeling ability and is mainly used in the fields of video timing signal processing, behavior recognition, etc. In this paper, Gaussian mixture HMM model is chosen to annotate the ethnic traditional sports movements at the frame level.

To solve the Baum-Welch algorithm’s tendency to fall into local optimal solutions, the K-means clustering algorithm is used to train the initial parameters for the observation sequences in the HMM model, and then the action features are extracted based on the spatial dependence of the joints, so as to obtain the feature matrices O={O1,⋯,Ot,⋯,OT} , Ot=[(Ot1)T,⋯(OtT)T,⋯,(OtW)W] of the action sequence J.

In the classifier training, the HMM model will use expectation maximization to compute the dependencies on the joint point rotations and the feature block timing dimensions. Where each class of action will be trained as an HMM model, denoted as {λ1,λ2,⋯,λc} , where c denotes the number of action classes. After obtaining the sequence of joint interest points J′ after (o′ is the corresponding feature matrix), the action is labeled by the HMM model for its action category based on the maximum likelihood probability. Then: (13)i*=argmaxi=1,2,⋯,LP(V/λi)

Meanwhile, in order to solve the temporal discontinuity of the traditional HMM model, this paper adopts the continuous Gaussian probability kernel function to calculate the probability density distribution of the observation value sequence. Namely: (14)bi(V)=∑m=1MGcmG(μim,Σ,V)

where G is the Gaussian probability kernel function.

A deep confidence network (DBN) consists of a stack of multiple restricted Boltzmann machines (RBMs), which is a probabilistic generative model containing one visible layer and multiple hidden layers [15]. The training process of DBN is divided into pre-training and fine-tuning. The RBMs are first pre-trained sequentially until all the RBMs constituting the DBN have completed the pre-training, and then the entire network of the DBN is fine-tuned using the gradient descent method. 1)

Pre-training. The purpose of pre-training is to reconstruct the input data. In pre-training, the output of each layer is used as the input of the lower network, i.e., the output of the first RBM is used as the input of the second RBM, and so on. The parameters of each layer of the RBM, i.e., connection weights and bias values, are learned by comparing them with the learning method of scatter until the last RBM is trained. The input of the lowest RBM is generally the observation data, also called the original input data, i.e., the ethnic traditional sports movement feature data extracted in this paper.

Given a training set {(h1k,y1),(h2k,y2),…,(hNk,yN)} , where N is the number of training samples, {h1k,h2k,…hNk} is the hidden layer vector of the top RBM, i.e., the input to the SoftMax layer, and assuming that w is the parameter of the SoftMax layer, the conditional probability p(y=j|hk),j=1,2,…,c of each class is: (15)[ p(yi=1|hi;w) p(yi=2|hi;w) M p(yi=c|hi;w)]=1∑j=1cewjThh[ ew1Thi ewi2hi M ewiThi]

where h_i ∈ Rⁿ⁺¹ denotes the hidden layer vector of the top RBM, w₁, w₂, …, w_c ∈ Rⁿ⁺¹ denotes the parameter connected to each output node, and 1∑j=1cewjThi is the normalization factor that makes the sum of all probabilities 1.

The cost function of SoftMax regression is similar to that of Logistic regression. Namely: (16)J(w)=−1N[∑i=1N∑j=1c1{yi=j}logp(yi=j|hik;w)]

Where N is the number of training samples, c is the number of labeled species, and 1{⋅} is an indicator function that takes the value 1 when the expression in curly brackets is true and 0 otherwise.

Its gradient is: (17)∇wjJ(w) = −1N∑i=1N[hi(1{yi=j}−p(yi=j|hik;w))] = −1N∑i=1N[hi[1{yi=j}−ewjThi∑j=1cewjThi)]

The gradient descent method is used to minimize the cost function, and the parameters are fine-tuned by back-propagation, at which time the weights of the top layer are connected to the parameter update formula: (18)wij=ηwij−ε∇wijJ(w)

where η is the momentum term, w_ij is the connection weight between the ird and jth nodes, and ε is the learning rate.

In order to avoid overfitting, a constraint term is usually added to the loss function as a regularization term, and the more widely used weight decay term is added, which has a weight decay term of λ2∑i∑i∑j(wij2)2 .

The cost function of adding the weight decay term in Eq. (16) is: (19)J^(w)=J(w)+λ2∑l∑i∑j(wjl)2

where λ denotes the coefficient of weight decay and l denotes layer l of the network. A regularization term is added to the cost function to avoid excessive weights, reduce overfitting and improve generalization.

In order to effectively enhance the recognition effect of ethnic traditional sports movements and promote the wide dissemination and innovative inheritance of ethnic traditional sports culture, this paper combines the DBN model with the HMM and establishes the DBN-HMM model for the recognition and classification of ethnic traditional sports movements. The structure of the DBN-HMM model is shown in Fig. 3, which is mainly used to estimate the a posteriori probability of the HMM to carry out the training and recognition and classification of ethnic traditional sports movements. training and recognition classification.

Figure 3.

The model structure of the DBN-HMM

After constructing the DBN-HMM model, it is necessary to initialize the network parameters as a way to iteratively train the RBM, and it is also necessary to tune the network, i.e., select the objective function for network parameter tuning. The specific operation is as follows:

First, the stochastic gradient descent method commonly used in DBN training is used for optimization, and the objective function adopts the cross-entropy between the reference state label and the predicted state distribution y(s) , whose mathematical expression is: (20)FCE=−∑u=1U∑n=1Nlogy(s)

where s is the state at moment t.

Then the output values of the DBN output layer nodes are used as inputs to the HMM, and the SoftMax regression model is used to compute the a posteriori probability of the appearance of the HMM state, and the output vector at moment t is set to be O_n. The expression of the output state distribution is given as: (21)yn(s)=P(s|On)=exp[an(s)]∑sexp[an(s)]

where an(s) is the activation probability of the output layer state s, i.e., the output value. There is a relational expression as: (22)logP(On|s)=logyn(s)−logP(s)

where P(s) is the prior probability of occurrence of state s in the training data.

Then the gradient expression between the objective function F_CE and the activation probability an(s) is: (23)∂FCE∂an(s)=−∂logyn(s)∂an(s)=yn(s)−δs:sn

where δs;sn is the Kronecker function satisfying. (24)δs;sn={ 1 s=sn 0 s≠sn

The feature vectors after preprocessing and feature extraction of the test ethnic traditional sports movements are inputted into the trained DBN-HMM model to get their state prediction labels. By comparing the predicted labels with the known test labels the recognition classification results of the ethnic traditional sports movements can be obtained.

Posture solving is the most basic part of motion capture technology, and choosing a suitable posture fusion algorithm is extremely important, this paper proposes complementary filtering algorithm for posture solving of ethnic traditional sports movements. In order to verify the effectiveness of the method in this paper, two groups of simple experiments are designed to carry out the performance of the pose solving algorithm. The first group is a static comparison test, in which the inertial sensor is placed horizontally on the desktop and kept at rest to collect the Euler angle data for 15s. The collected data is divided into three groups, the first group is the Euler angle solved by data fusion of the nine-axis IMU using Kalman filtering algorithm. The second group is the Euler angles solved by data fusion of nine-axis IMU using Mahony algorithm, and the third group is the Euler angles solved by data fusion of nine-axis IMU using complementary filtering algorithm. Fig. 4 shows the comparison results of attitude solved static data, where Fig. 4(a)~(c) shows the comparison results of roll angle, pitch angle and yaw angle, respectively.

Figure 4.

The comparison of static data

The roll angle, pitch angle and yaw angle data obtained by the nine-axis IMU attitude solution algorithm based on the complementary filtering algorithm are smoother, and their fluctuation ranges are smaller, which indicates that this paper’s method can better solve the attitude data when performing the motion capture of ethnic traditional sports. On the other hand, the attitude solving results based on Mahony algorithm and Kalman filtering algorithm show a large range of fluctuation. The smoother the posture solving data are, the easier it is to get the standard deviation and extreme deviation of the captured data by using the tools of MATLAB, so as to compare the advantages and disadvantages of different posture solving methods more comprehensively.

The second group is a dynamic comparison test. A digital angle ruler was utilized to mimic the human joint movement thus comparing the dynamic measurements of the three methods one end of the ruler was fixed horizontally at the wall end of the table, and the other end was used as the movable axis for rotational movement. The IMU was fixed horizontally (where the Z-axis was horizontally perpendicular to the table) on the movable axis to measure the Euler angle data. First, the initial state of the ruler was kept at 0°, then the active axis was rotated to the right at a uniform speed to 120°, and the attitude solution was performed simultaneously using the three algorithms, and the comparative results of the yaw angle data solution are shown in Fig. 5.

Figure 5.

The comparison results of the yaw data solution

It can be clearly seen that the trend of the attitude angle transformations solved by the three algorithms is basically the same from 0° to 120°. By careful observation, the attitude angle data curve using the complementary filtering algorithm is slightly smoother compared to that of the Mahony algorithm and the Kalman algorithm. In addition, when data fusion is performed with the three algorithms, the response time of Mahony algorithm and Kalman algorithm is 0.12ms and 0.09ms, respectively, while the response time of this paper’s method is only 0.02ms, which can be seen that this paper’s method’s response time is significantly lower than that of the other two algorithms. Considering that multiple node sensors need to be solved at the same time in the application of ethnic traditional sports motion capture, and the high response time is easy to cause data loss, the complementary filtering algorithm used in this paper for the fusion gesture solving of sports motion data has a high feasibility.

The experiment selected eight ethnic traditional sports athletes for sports motion capture test, and collected the data of ethnic traditional sports movements of each test subject. The experiment was set up with seven movement positions: left and right arm swing (X1), up and down arm swing (X2), push and pull movement (X3), chest expansion movement (X4), stretching movement (X5), jumping posture (X6), and high-five posture (X7), and the test subjects were required to repeat the execution of each movement for 10 times, resulting in 80 samples of movement data. 80% and 20% of the experimental samples are used as training samples and test samples, respectively, and the experimental samples are prepared for use after the implementation of data preprocessing and normalization, and at the same time, the same-condition comparative test is introduced to the body area network (BAN)-based, monocular vision-based, and OpenPose-based motion capture technologies, in order to highlight the advantages of the method of this paper in the aspect of motion capture.

Athletes demonstrated national traditional sports movements, and four methods were used to capture the movement posture data of nine body parts of the test subjects, namely, head, neck, left shoulder, right shoulder, waist, left wrist, right wrist, left foot, right foot, and the human body joints data captured by Kinect were regarded as the standard data, and the error results of the four methods of movement capture were counted as shown in Figure 6.

Figure 6.

Error results of action capture

The error curve of this paper’s method is located at the bottom of the image, indicating that the method has the smallest error in the measurement of ethnic traditional sports movements, and the measurement errors of the nine parts are between 0.243mm and 0.891mm, and the fluctuation of the error data is small, and the measurement results are more stable. Comparatively speaking, the error of monocular vision-based motion capture technology continues to rise, and the highest error can reach 2.935mm, and the error fluctuation interval of body domain network-based motion capture technology is [1.272mm,2.463mm], and the overall fluctuation is large. And although the OpenPose-based motion recognition technology does not have much error fluctuation, the measurement error value of this method is much higher than that of this paper’s method, and the motion error capture effect is not ideal. It can be seen that this paper’s method demonstrates the best measurement effect of ethnic traditional sports movements.

In order to further illustrate the efficiency of this paper’s method in performing ethnic traditional sports movement capture, further expanding the data of ethnic traditional sports movements, selecting 10~60 ethnic traditional sports movements, respectively, and continuing to test the efficiency of this paper’s method and the above three methods in performing ethnic traditional sports movement capture. Figure 7 shows the test results.

Figure 7.

The comparison of action capture efficiency

When carrying out ethnic traditional sports movement capture, the shorter the capture time, the better the capture effect of the capture method, and the longer the capture time, the worse the capture effect of the capture method. As can be seen from the figure, with the increase of the number of ethnic traditional sports movements, the capture time tested by the four methods showed different degrees of upward trend. However, the capture time tested by this paper’s method is the lowest among the three methods, even when the number of ethnic traditional sports movements is 60, the capture time is only 6.28 s. This is mainly due to the fact that this paper’s method implements gesture solving on the data in time after obtaining the data of the ethnic traditional sports movements, so that the method is more efficient in capturing the ethnic traditional sports movements.

In order to test the effectiveness of the recognition model of ethnic traditional sports movements based on DBN-HMM model, a dataset is established based on 80 data samples of ethnic traditional sports movements in section 4.1.2. The HMM model is used to annotate the ethnic traditional sports movements at the frame level, and then the DBN model is used to recognize and classify the ethnic traditional sports movements. 70% of the dataset is selected for model training, and the remaining 30% is used as the test set for model testing. Fig. 8 shows the recognition results of different ethnic traditional sports actions, where Fig. 8(a) and (b) show the recognition confusion matrix of static features and static+dynamic features, respectively.

Figure 8.

Identification of traditional national sports actions

The recognition results are represented in the form of confusion matrix plots, with the diagonal line being the correct recognition rate of the corresponding category and the rest being the recognition rate of the category judged to be incorrect. Comparison of the two confusion matrix plots shows that the mean accuracy of the recognition results using only static features is 0.72, which is significantly lower than that of the recognition results combining static and dynamic features of 0.92. It indicates that the addition of dynamic features can effectively improve the recognition effect, and the combination of dynamic and static features can enhance the characterization ability of the traditional national sports movements.

In order to further illustrate the efficiency of the method in this paper, the recognition times of seven different categories of ethnic traditional sports movements in the dataset are counted, and PSO-BPNN, GMM-SVM, and DNN-HMM models are selected as comparisons to validate the differences in the recognition times of different movements with different algorithms. Figure 9 shows the average recognition time comparison results of the four recognition models.

Figure 9.

The average recognition time comparison results of the model

It can be seen that the DBN-HMM model proposed in this paper has the shortest recognition time for ethnic traditional sports movements, and the average recognition time for the seven movement categories is 24.77 s. In this paper, we use the HMM model to first perform the frame-level labeling for ethnic traditional sports movements, and then use the DBN model to obtain the a posteriori probability of the HMM, which effectively reduces the complexity of the model, and enhances the identification of ethnic traditional sports movements. The modeling complexity is effectively reduced, and the recognition efficiency of ethnic traditional sports movements is improved. The model in this paper is significantly more efficient than the traditional SVM and BPNN methods in recognizing and classifying traditional ethnic sports movements, which can complete the recognition of traditional ethnic sports movements, provide support for the promotion of the dissemination of traditional ethnic sports culture, and meet the practical application requirements.

In order to more intuitively observe the distribution of the input ethnic traditional sports action sequences and the differences in the results of different methods for action recognition experiments, this paper chooses the better-performing GMM-SVM as a comparison, and selects X1 and X2 from seven action types for comparison. Fig. 10 shows the results of comparing the action recognition sequences of the two models for the two action types, where Fig. 10(a) and (b) show the input sequences of sports actions X1 and X2, Fig. 10(c) and (d) show the recognition sequences of the GMM-HMM model, and Fig. 10(e) and (f) show the recognition sequences of the DBN-HMM model. As can be seen from the figures, compared with the GMM-HMM model, the recognition results of the 2 ethnic traditional sports movement recognition experiments using the DBN-HMM model have a much smaller gap between the recognition results and the actual input movement sequences, and the recognition accuracy is higher, which verifies the feasibility and superiority of the DBN-HMM model in the recognition scheme of the ethnic traditional sports movements from another perspective.

Figure 10.

Action identification sequence comparison results

Cultural self-consciousness is built on the search for and inheritance of the “root”, which is a profound understanding of the status and role of culture, a correct grasp of the laws of cultural development, and an initiative to take responsibility for the development of cultural history. Only when the main body of inheritance consciously assumes the social responsibility of inheritance and constantly enhances the sense of cultural identity, will it practice the inheritance of traditional national sports non-legacy from the bottom of its heart, and ultimately form a virtuous circle of traditional national sports non-legacy protection and inheritance mode. Using the identified national traditional sports movements, integrating them into the physical education curriculum of colleges and universities, helping students to experience the national traditional sports culture more intuitively, enhancing the cultural consciousness and sense of cultural identity of the inheriting body, and boosting the popularization and inheritance of traditional sports culture.

The inheritance of national traditional sports culture requires a systematic teaching and training system to standardize every technical movement and refine the internal laws of national traditional sports culture. For example, in traditional martial arts and folk dances, each technical movement of step, hand and eye should be taught precisely. With the help of technology to accurately identify the national traditional sports movements, invite the inheritors to carry out the standardized guidance of technical movements, in-depth research and build a systematic and standardized technical system, so that the national traditional sports projects form an organic and unified inheritance technology system. We should also pay attention to the theoretical and practical research on the national traditional sports culture, study its spiritual connotation and value of the times, and in the inheritance process, we should not only inherit the external skills, routines and forms, but also inherit the spiritual connotation inherent in the non-legacy of sports. It is necessary to make full use of the resource advantages of colleges and universities to create a platform for displaying traditional national sports culture, so as to form an inheritance system integrating culture and technology.